tglkmeans 0.6.3

tglkmeans 0.6.3

• Fix: metric = "spearman" ignored missing values. Ranking tested the wrong
  missing-value sentinel, so NAs were ranked as the largest value and included
  in the rank correlation instead of being dropped. Spearman now excludes missing
  values pairwise, matching euclid/pearson. Clustering results for Spearman on
  data with NAs change (and are now correct); results on complete data are
  unchanged.
• Fix: predict_tgl_kmeans() with metric = "euclid" used a plain Euclidean
  distance, which disagreed with the training metric sqrt(sum_sq) / n when a
  cluster center had a missing dimension. Prediction now reproduces the training
  distance exactly. Predictions on data whose centers have no missing dimensions
  are unchanged.
• Performance: removed the dense k x n per-thread vote matrix in the
  reassignment step. Memory and per-iteration work no longer scale with the
  number of clusters; cluster assignments are unchanged.
• Performance: metric = "spearman" no longer computes a discarded p-value on
  every point-to-center comparison, and uses a contiguous sort buffer instead
  of a linked list. Results are unchanged.
• Performance: predict_tgl_kmeans() uses smaller internal chunks, cutting
  redundant distance/correlation computation on large inputs. Results are
  unchanged.
• Fix: hclust_intra_clusters = TRUE returned a scrambled within-cluster
  ordering. The order/intra_clust_order columns now follow the hclust
  dendrogram leaf order as documented.
• Removed unused internal code (reduce_coclust/reduce_num_trials and dead
  rank-sum / incomplete-beta helpers).

tglkmeans 0.6.2

• Fix: predict_tgl_kmeans() crashed with 'from' contains NAs / NAs introduced by coercion to integer range on inputs of ~46K rows or more. The one-shot as.matrix(tgs_dist(.)) overflowed integer indexing inside stats:::as.matrix.dist. The prediction now processes observations in chunks (#21).

tglkmeans 0.6.1

• Added predict_tgl_kmeans() function to assign new observations to existing k-means cluster centers (#5).
• Exported match_clusters() and test_clustering() functions.
• Auto-detect character/factor first column as ID column.
• Parallelized k-means initialization.
• Fix: Pearson distance was not negated, causing incorrect cluster assignments when using metric = "pearson".
• Fix: memory leak in k-means core.
• Fix: k-means seeding crash when k is large relative to data size.
• Fix: race condition in parallel workers.
• Fix: package failed to load on machines where detectCores() returns NA.
• Fix: downsample_matrix used identical random seed for all columns.
• Removed plyr dependency. Moved ggplot2 from Imports to Suggests.